Apple's SHARP Gaussian splat model runs entirely in-browser via ONNX

A developer has ported Apple’s SHARP model to run client-side in the browser using ONNX Runtime Web, turning a single uploaded image into a downloadable Gaussian splat .ply file with no server-side inference. The pipeline uses a React/TypeScript UI, a Web Worker for ONNX inference, and browser-side postprocessing that converts NDC coordinates to metric gaussians before writing the PLY output. A WebGPU-capable desktop browser is effectively required, and the exported model ships as a ~2.4 GB ONNX graph plus a separate .onnx.data weights sidecar that must be co-located.

The project is a thin shim around upstream Apple tooling rather than a reimplementation. Users export their own checkpoint via a provided Python script that wraps Apple’s repo, or rely on a hosted model by default since uploading only the .onnx graph without its weight sidecar will fail. Apple ships SHARP code and model weights under separate licenses, with the weights restricted to research use, which constrains any downstream productization.

The more interesting signal is architectural: a multi-gigabyte vision model executing entirely in WASM/WebGPU with no inference backend means zero per-request GPU cost and no image leaving the device. The tradeoff is brittle compatibility, heavy first-load times, and memory pressure that pushes the experience toward enthusiast hardware. It is labeled experimental, and that label is honest.